Clinical validation of a morphometric artificial intelligence system for detecting osteoporotic vertebral compression fractures on plain radiographs

Osteoporotic vertebral compression fractures (OVCFs) are a major cause of morbidity and mortality in aging populations, and plain radiography—the most available modality for initial assessment—is operator-dependent. This study evaluated the diagnostic performance and clinical utility of a deep learning-based morphometric artificial intelligence (AI) model for detecting OVCFs and quantifying vertebral compression rates on plain radiographs. A total of 100 patients (780 vertebral bodies; 685 normal, 95 fractured) with suspected vertebral compression fractures were retrospectively analyzed. All patients underwent spine radiographs followed by computed tomography or magnetic resonance imaging within 5 weeks and bone mineral density assessment within 1 year. The AI model calculated vertebral compression rates and classified fractures using a 20% compression threshold. AI performance was compared with blinded assessments by two orthopedic surgeons, one orthopedic resident, and one general physician. Metrics were reported with cluster-adjusted 95% confidence intervals (patient-level bootstrap) and compared using McNemar’s test; balanced accuracy, F1 score, a patient-level analysis, and compression-rate agreement (intraclass correlation) were also assessed. The AI model demonstrated high specificity (96.6%) and negative predictive value (95.4%). Its sensitivity, however, was lower than that of physicians (66.3% vs. 75.8%–86.3%), with lower balanced accuracy (81.5% vs. 86.5%–92.3%) reflecting the low fracture prevalence (12.2%); on McNemar testing, AI accuracy and sensitivity were significantly lower than the surgeons’ and resident’s but not the general physician’s, whereas specificity did not differ. Patient-level sensitivity was 84.3% (negative predictive value 70.3%). AI sensitivity rose with severity, from approximately 48% for sub-threshold (< 20%) fractures to 100% for Genant grade 3. AI-derived compression rates correlated with physicians’ measurements ( r = 0.594–0.873; intraclass correlation 0.534–0.850; all p < 0.01), and the area under the ROC curve was 0.8095 for the AI model versus 0.8651–0.9228 for physicians. The morphometric AI model showed high specificity and meaningful correlations with physician-derived measurements but did not outperform physician evaluators, and its sensitivity (66.3%) was insufficient for stand-alone use. It is therefore not suitable for independent diagnostic, rule-out, or triage use; rather, it may act as an adjunctive aid that supports physician interpretation, chiefly by providing objective, reproducible quantification of vertebral compression.